An edge-computing-based industrial reverse osmosis system intelligent operation and maintenance method

Abstract

The application relates to the field of operation and maintenance analysis, and discloses an intelligent operation and maintenance method for an industrial reverse osmosis system based on edge computing, which realizes real-time collection of reverse osmosis system operation data through an edge computing node, forms an operation parameter data set after pretreatment, inputs the data set into a first prediction model to predict short-term change values of key parameters, selects a verification mode according to real-time operation conditions of the system to logically verify the first prediction result, inputs the verified result into a second prediction model to obtain a comprehensive health index, obtains an operation and maintenance decision index by fusing the double prediction results, generates differentiated operation and maintenance guidance decisions of continuous monitoring, optimized adjustment, preventive maintenance or immediate intervention, and outputs a targeted fault diagnosis result after triggering a parameter feedback analysis mechanism to locate abnormal parameters and replacing the abnormal parameters with historical statistical values for re-verification if the verification fails. The application solves problems such as lagging response of traditional operation and maintenance and inaccurate prediction, improves system operation stability, and reduces operation and maintenance cost.

Description

Technical Field

[0001] This invention relates to the field of operation and maintenance analysis, and specifically to an intelligent operation and maintenance method for industrial reverse osmosis systems based on edge computing. Background Technology

[0002] Industrial reverse osmosis systems are widely used in various industrial fields such as water treatment, petrochemicals, power, and electronics. Their core function is to achieve material separation and purification through the selective permeability of semi-permeable membranes, making them key equipment for water resource recycling and process fluid purification in industrial production. With the increasing scale and automation of industrial production, higher demands are placed on the operational stability, product water quality, and maintenance efficiency of reverse osmosis systems.

[0003] Currently, the operation and maintenance methods for industrial reverse osmosis systems mainly include two categories: traditional manual operation and maintenance and preliminary intelligent operation and maintenance. Traditional manual operation and maintenance relies on operators to conduct regular inspections, record operating parameters, and judge the system status and perform maintenance based on experience. This method has problems such as poor real-time performance, high labor intensity, and delayed fault warning. It is often only discovered after obvious faults occur in the system (such as a sudden drop in permeate flow, excessive desalination rate, and severe membrane fouling). This not only leads to substandard permeate quality and production interruption, but may also cause irreversible damage to membrane elements due to the expansion of the fault, significantly increasing operation and maintenance costs.

[0004] Existing preliminary intelligent operation and maintenance solutions mostly rely on cloud computing architectures, transmitting collected operational data to the cloud for processing and prediction. However, industrial sites generate massive amounts of data with high real-time requirements. Cloud computing suffers from drawbacks such as data transmission latency, high network bandwidth consumption, and untimely edge response, making it difficult to meet the needs of real-time monitoring and rapid decision-making for key parameters of reverse osmosis systems. Furthermore, existing prediction models often use single validation rules to verify prediction results, which cannot adapt to different operating conditions such as steady-state operation, gradual performance changes, and sudden parameter changes, leading to insufficient accuracy and a tendency for misjudgments or omissions. Therefore, there is an urgent need for an intelligent operation and maintenance method for industrial reverse osmosis systems based on edge computing. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent operation and maintenance method for industrial reverse osmosis systems based on edge computing, thereby solving the above-mentioned technical problems.

[0006] The objective of this invention can be achieved through the following technical solutions: An intelligent operation and maintenance method for industrial reverse osmosis systems based on edge computing includes: The operating data of the reverse osmosis system is collected and preprocessed in real time at the edge computing node, and time-domain and frequency-domain features are extracted to form an operating parameter dataset. The operating parameter dataset is input into the first prediction model deployed at the edge to predict the short-term changes of key operating parameters within a preset time period, which is used as the first prediction result; the key operating parameters include at least the water production flow rate, desalination rate, and inter-stage pressure difference; The first prediction result is logically verified, and the verification mode used in the logical verification is dynamically selected according to the real-time operating conditions of the system. If the logic verification passes, the first prediction result is input into the second prediction model to evaluate the system's comprehensive health index as the second prediction result. The first prediction result and the second prediction result are then fused based on composite operations to generate operation and maintenance guidance decisions. If the logic verification fails, the parameter feedback analysis mechanism is triggered to locate the abnormal operating parameters. The historical statistical values ​​of the abnormal operating parameters are then used as substitutes before the logic verification is performed again. If the logic verification after substitution passes, the device fault diagnosis result for the abnormal operating parameters will be output.

[0007] As a preferred embodiment, the trigger parameter feedback analysis mechanism for locating abnormal operating parameters includes: For each operating parameter, calculate the average rate of change of the cumulative physical quantity of that operating parameter within a preset time window relative to the window length, as the cumulative change trend of that operating parameter; The average rate of change is compared with the baseline range of the corresponding parameter established based on historical normal operation data. If it exceeds the baseline range, the operating parameter is determined to be an abnormal operating parameter.

[0008] As a preferred option, the verification mode is dynamically selected based on the real-time operating conditions of the system, including: Based on the stability of key operating parameters, select the first set of verification rules for steady-state verification; Based on the gradual trend of system performance indicators or a preset period, select a second set of verification rules for in-depth performance verification. Based on the sudden changes or exceeding of limits in the operating parameters, a third set of verification rules is selected for anomaly and boundary checks.

[0009] As a preferred embodiment, the verification of the first verification rule set simultaneously satisfies: The difference between the real-time collected influent flow rate and the sum of the product water flow rate and concentrate flow rate is less than the first allowable error threshold. The real-time ratio of inlet water pressure to product water flow rate is within the normal reference range determined based on historical steady-state data of the system.

[0010] As a preferred embodiment, the verification of the second verification rule set simultaneously satisfies: The system recovery rate is based on real-time traffic calculation, and the standard deviation of the system recovery rate is less than the preset stability threshold. The absolute deviation between the system desalination rate predicted based on operating parameters and the desalination rate calculated based on real-time water quality measurements is less than a preset consistency threshold. The rate of change of the moving average of the inter-segment pressure difference within the preset time period is less than the preset pollution growth threshold.

[0011] As a preferred embodiment, the verification of the third verification rule set simultaneously satisfies: The real-time calculated system recovery rate is ≤ the maximum allowable recovery rate specified for the membrane element, and the real-time inlet water pressure is ≤ the system safety pressure limit; When the change in influent conductivity exceeds the abrupt change threshold, the predicted change in product water conductivity is consistent with the theoretical response direction determined based on membrane desalination characteristics. After a step change in the influent flow rate, the response time for the predicted product flow rate to reach a steady state is within the theoretical time range calculated based on the system's hydraulic volume.

[0012] As a preferred embodiment, the first prediction model is a lightweight long short-term memory network model; the second prediction model is a Transformer encoder model, used to output a comprehensive health index that characterizes the overall health status of the system.

[0013] As a preferred embodiment, the fusion of the first and second prediction results based on composite operations is achieved by calculating a multi-dimensional operation and maintenance decision index, specifically including: From the first prediction result, extract the normalized vector ΔV=[δ1,δ2,...,δn] of the short-term change values ​​of the key parameters; The comprehensive health index H is obtained from the second prediction result, wherein a decrease in the H value indicates a deterioration in health status; The operation and maintenance decision index A is calculated using the following composite formula: ; Where |ΔV| represents the norm of the normalized vector of the short-term change values ​​of the key parameters extracted from the first prediction result; The range of H is [0,1], and the lower the H value, the worse the health status. This represents a preset health status threshold, used to divide different decision intervals; The preset weighting coefficients, The preset threshold for the magnitude of change; This is the Sigmoid function, used to provide a smooth decision transition; The operation and maintenance decision index A is compared with a preset decision threshold to generate a corresponding operation and maintenance guidance decision.

[0014] As a preferred approach, generating operation and maintenance guidance decisions specifically includes: At least three ordered decision thresholds are preset: routine monitoring threshold K1, early warning and maintenance threshold K2, and emergency response threshold K3, where K1 < K2 < K3; The calculated operation and maintenance decision index A is compared with the decision threshold, and a corresponding preset decision template is matched according to the interval: If A≤K1, then the continuous monitoring decision template is matched; If K1 < A ≤ K2, then match and optimize the decision template. If K2 < A ≤ K3, then the preventive maintenance decision template is matched; If A > K3, then the immediate intervention decision template is matched.

[0015] As a preferred solution, if the logic verification still fails after the substitution, the parameter feedback analysis mechanism will be re-triggered, and a secondary fault location will be executed: The preset time window is shortened to a second duration, and the average rate of change of all operating parameters within the shortened window is calculated. Simultaneously, collect and analyze the system's hardware status data; Based on the average rate of change and the hardware status data, the fault type is determined to be either a multi-parameter collaborative anomaly or a hardware fault, and the corresponding secondary fault diagnosis result is output.

[0016] The beneficial effects of this invention are: (1) This invention uses edge computing nodes to collect and preprocess running data in real time, and deploys data processing and models on the edge side, avoiding the data transmission delay and bandwidth occupation problems of traditional cloud computing, realizing local real-time processing of running data, and timely capturing subtle changes in system parameters, providing timely and reliable data support for subsequent prediction and decision-making, and effectively improving the real-time performance of operation and maintenance response.

[0017] (2) The present invention dynamically selects three types of verification rule sets, namely steady-state verification, performance depth verification, and anomaly and boundary verification, based on the real-time operating conditions of the system. This overcomes the defect of the single verification mode in the prior art and ensures the effectiveness and accuracy of the first prediction result under different operating conditions. At the same time, the lightweight long short-term memory network LSTM model is used to accurately predict short-term parameter changes. Combined with the Transformer encoder model to evaluate the overall health index of the system, the invention achieves dual coverage of short-term parameter change prediction and overall health status evaluation. The prediction dimension is more comprehensive, and the misjudgment rate and missed judgment rate are greatly reduced.

[0018] (3) The present invention sets up a parameter feedback analysis mechanism. By calculating the average rate of change of the operating parameters within a preset time window and comparing it with the benchmark range established by historical normal data, abnormal operating parameters can be quickly and accurately located. At the same time, by replacing the historical statistical values ​​of abnormal parameters with re-verification, the interruption of the entire operation and maintenance process due to a single abnormal parameter is avoided, and the accuracy of fault diagnosis is improved. Attached Figure Description

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] Figure 1 This is a diagram illustrating the method steps of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 As shown, this invention is an intelligent operation and maintenance method for industrial reverse osmosis systems based on edge computing, comprising: The operating data of the reverse osmosis system is collected and preprocessed in real time at the edge computing node, and time-domain and frequency-domain features are extracted to form an operating parameter dataset. For example, the edge computing node collects data from each sensor at a frequency of 1 time / second. The raw data collected at a certain moment is as follows: influent flow rate: 48.5 m³ / h, product water flow rate: 35.2 m³ / h, concentrate flow rate: 13.1 m³ / h; influent pressure: 1.2 MPa, inter-stage pressure difference: 0.35 MPa; influent conductivity: 850 μS / cm, product water conductivity: 15 μS / cm; influent temperature: 25℃.

[0023] The preprocessing process is as follows: No outliers (all parameters are within the sensor measurement range and the system's normal operating range), no missing values, directly retained; Z-score normalization is used to convert all parameters into dimensionless values; Time-domain characteristics: Calculate the mean (e.g., mean influent flow rate 48.3 m³ / h), variance (0.21), and peak value (49.1 m³ / h) of parameters such as influent flow rate and product flow rate over the past 5 minutes. 3 / h); Frequency domain features: Fast Fourier Transform (FFT) is performed on the inlet pressure and inter-section pressure difference data to extract the frequency spectrum peak (0.02Hz) and the main frequency band energy ratio (85%); finally, a 32-dimensional operating parameter dataset is formed and input into the first prediction model.

[0024] The runtime parameter dataset is input into the first prediction model deployed at the edge. The first prediction model is a lightweight long short-term memory network model. Model structure: Input layer: 32 dimensions (corresponding to 32-dimensional time domain + frequency domain features, including the mean, variance, peak value, and frequency spectrum peak value of each operating parameter); Hidden layers: 2 lightweight LSTM layers, with 64 neurons in the first layer and 32 neurons in the second layer, and a Dropout layer (dropout rate=0.2) is used to prevent overfitting; Output layer: 3-dimensional (short-term predicted values ​​of corresponding permeate flow rate change rate, desalination rate change rate, and inter-stage pressure difference change rate); Activation functions: The hidden layer uses the ReLU function, and the output layer uses the Linear function.

[0025] A lightweight LSTM structure is adopted, which reduces the number of neurons (64→32) and simplifies the network layers, thereby reducing the computational cost of the model while ensuring prediction accuracy and adapting to the hardware resources of edge computing nodes. The introduction of the Dropout layer effectively avoids model overfitting and improves the model's generalization ability under different operating conditions. Compared with the traditional tanh function, the ReLU activation function alleviates the gradient vanishing problem and accelerates the model training convergence speed.

[0026] Training process and training data: The reverse osmosis system was continuously monitored for three months, resulting in approximately 86,400 samples (1 time / second × 3,600 seconds × 24 hours × 30 days). These samples were divided into a training set (60,480 samples) and a validation set (25,920 samples) in a 7:3 ratio. Optimizer: Adam optimizer, with an initial learning rate of 0.001 and a learning rate decay strategy (decaying to 0.9 every 100 rounds). Loss function: Mean Squared Error (MSE), target loss value ≤ 0.005; Training rounds: 300 rounds, batch size=128, after training the model inference time ≤50ms / inference, meeting the real-time requirements of edge computing.

[0027] The model is trained using large-scale historical normal operation data (86,400 sets of samples) to ensure that the model learns the characteristics and patterns of stable system operation; the learning rate decay strategy balances the convergence speed in the early stage of model training and the training accuracy in the later stage; the inference time of 50ms / inference meets the real-time prediction requirements of the edge side and can quickly output the short-term change values ​​of key parameters, thus saving time for subsequent logic verification and operation and maintenance decisions.

[0028] Predict the short-term changes of key operating parameters within a preset time period as the first prediction result; key operating parameters include at least permeable flow rate, desalination rate, and inter-stage pressure difference; Example: Input a 32-dimensional dataset into a lightweight LSTM model to predict the short-term changes in key operating parameters within the next hour (preset time period), and obtain the first prediction result: Production water flow rate change rate: +2.1% (meaning the production water flow rate is expected to increase to 35.2 × (1 + 2.1%) = 35.9 m³ in the next hour). 3 / h); Desalination rate change rate: -0.8% (Current desalination rate = (1-15 / 850)×100%=98.24%, expected to decrease to =97.46%). Inter-segment pressure difference change rate: +5.3% (expected to increase to 0.35×(1+5.3%)=0.37MPa).

[0029] The lightweight LSTM model accurately predicts the short-term change trends of key parameters within the next hour, providing a forward-looking basis for operation and maintenance decisions. Key parameters (product water flow rate, desalination rate, and inter-stage pressure difference) are directly related to the core performance of the system and the status of equipment. Accurate prediction of their change trends can help operation and maintenance personnel identify potential risks in advance and avoid passively responding to failures.

[0030] The first prediction result is logically verified, and the verification mode used in the logical verification is dynamically selected according to the real-time operating conditions of the system. The verification mode is dynamically selected based on the real-time operating conditions of the system, including: Based on the stability of key operating parameters, a first set of verification rules is selected for steady-state verification; the verification of the first set of verification rules simultaneously satisfies: The difference between the real-time collected influent flow rate and the sum of the product flow rate and concentrate flow rate is less than the first permissible error threshold. The first permissible error threshold originates from the overall accuracy error of the flow metering equipment. This value is obtained through a calibration certificate and represents the allowable instrument error range when verifying the law of conservation of mass: influent flow rate = product flow rate + concentrate flow rate.

[0031] The real-time ratio of influent pressure to product water flow rate is within the normal reference range determined based on the system's historical steady-state data. By analyzing the system's historical clean and stable operation data over several months, the distribution of this ratio on high-frequency data points (e.g., using the 5%-95th percentile range) is statistically analyzed to form an empirical normal range.

[0032] Example: First, determine the real-time operating condition of the system: the fluctuation range of key operating parameters (product water flow rate, desalination rate, and inter-stage pressure difference) in the last 10 minutes is ≤3%, indicating a steady-state operation. Therefore, the first set of verification rules is selected for logical verification.

[0033] First verification rule set verification process: Verification Rule 1: The percentage difference between the influent flow rate and the sum of the product flow rate and concentrate flow rate. Calculation: 48.5 - (35.2 + 13.1) = 0.2 m³ 3 / h, the difference percentage = 0.2 / 48.5 = 0.41% < the first allowable error threshold (5%), which meets the rule; Verification Rule 2: Real-time ratio of inlet water pressure to product water flow rate. Calculation: 1.2 MPa / 35.2 m 3 / h=0.0341MPa·h / m 3 ; Historical steady-state data normal reference range: Based on the statistical data of steady-state operation over the past 3 months, the 95% confidence interval for this ratio is [0.032, 0.036] MPa·h / m 3 The current ratio is within the range, satisfying the rule. Conclusion: The first prediction result passes logical verification.

[0034] The first set of verification rules is dynamically selected based on the real-time operating conditions (steady state) of the system, which achieves accurate matching between the verification mode and the operating state and avoids the problem of insufficient adaptability of a single verification rule to different operating conditions. The two verification rules verify the prediction results from the perspectives of flow balance and pressure-flow correlation, respectively, ensuring the rationality and reliability of the first prediction result and laying the foundation for subsequent health index assessment and operation and maintenance decisions.

[0035] Based on the gradual trend or preset period of system performance indicators, a second set of verification rules is selected for in-depth performance verification; the verification of the second set of verification rules simultaneously satisfies: The system recovery rate is based on real-time traffic calculation, and the standard deviation of the system recovery rate is less than the preset stability threshold. The absolute deviation between the system desalination rate predicted based on operating parameters and the desalination rate calculated based on real-time water quality measurements is less than a preset consistency threshold. Stability threshold and pollution growth threshold: These are set by statistically summarizing the values ​​of the standard deviation of system recovery rate fluctuation and the rate of change of inter-segment pressure difference in historical data during the normal slow accumulation phase of pollution (e.g., taking the maximum value or average value plus a certain margin), and are used to distinguish between normal fluctuations and abnormal accelerated pollution.

[0036] Consistency threshold: Used to compare the predicted desalination rate with the measured desalination rate. This threshold mainly considers the measurement accuracy of the online conductivity meter and the reasonable error of the prediction model itself, and is usually set as the square root of the sum of the squares of the errors of the two.

[0037] The rate of change of the moving average of the inter-segment pressure difference within a preset time period is less than the preset fouling growth threshold. Performance depth verification is triggered at preset intervals to periodically assess the long-term operating performance of the system and promptly identify gradual problems (such as slow growth of membrane fouling). Three rules verify the system's core performance indicators from the perspectives of recovery rate stability, desalination rate prediction accuracy, and membrane fouling growth rate, ensuring long-term stable operation.

[0038] Based on the sudden changes or exceeding of limits in operating parameters, a third set of verification rules is selected for anomaly and boundary checks. Sudden change: refers to a change in operating parameters (such as influent conductivity, influent flow rate) exceeding a sudden change threshold within a timescale much shorter than the system hydraulic residence time or process response time. The sudden change threshold is typically set based on a combination of sensor accuracy, normal fluctuation amplitude, and process sensitivity. In this invention, "much shorter than the system hydraulic residence time" specifically means shorter than 1 / 5 of the minimum hydraulic residence time calculated based on system design parameters. For example, for the system in this embodiment, the minimum hydraulic residence time is 10 minutes, then the timescale refers to less than 2 minutes.

[0039] Exceeding limits: refers to operating parameters (such as pressure, recovery rate) exceeding the hard limits specified in the safety data sheets (such as the maximum permissible recovery rate of membrane elements, the pressure resistance limits of pumps and pipelines) or the safety margins of the process design provided by the equipment manufacturer.

[0040] The verification of the third verification rule set simultaneously satisfies: The real-time calculated system recovery rate is less than or equal to the maximum allowable recovery rate specified by the membrane element, and the real-time inlet water pressure is less than or equal to the system safety pressure limit. The maximum allowable recovery rate and the system safety pressure limit are directly derived from the technical manual of the membrane element manufacturer and the rated working pressure of the system pressure vessel, pipeline, and pump, and are safety red lines that cannot be crossed. When the change in influent conductivity exceeds the abrupt change threshold, the predicted change in product water conductivity is consistent with the theoretical response direction determined based on membrane desalination characteristics. After a step change in the influent flow rate, the response time for the predicted permeate flow rate to reach steady state falls within the theoretical time range calculated based on the system's hydraulic volume. The theoretical response direction and theoretical time range: Based on the mass transfer kinetics theory of membrane separation using existing technology and the hydraulic residence time calculated from the system's pipeline volume and flow rate, this is the foundation for determining whether the system's dynamic response is normal; further details will not be elaborated upon here.

[0041] Example: At a certain moment, the conductivity of the influent suddenly changes, that is, it rises from 850 μS / cm to 1400 μS / cm within 1 hour. The change amount = 550 μS / cm > the sudden change threshold of 50 μS / cm. The system condition is judged to be a parameter sudden change, and the third verification rule set is selected for verification: Verification rule 1: System recovery rate = 35.2 / 48.5 × 100% = 72.6% ≤ 75% (maximum allowable recovery rate), inlet water pressure = 1.2MPa ≤ 1.6MPa (safe pressure limit), the rule is met; Verification Rule 2: As the influent conductivity increases, theoretically the product water conductivity should increase in tandem with the membrane desalination characteristics. However, the product water conductivity output by the first prediction model decreases, which is inconsistent with the theoretical response direction and does not meet the rule. Verification Rule 3: No step change in influent flow rate occurred, no verification is required.

[0042] When the system experiences a sudden change in parameters, the third set of verification rules is dynamically selected to specifically verify the rationality of the prediction after the parameter change. Rule 2 verifies the consistency between the predicted change direction and the theoretical response direction of the membrane desalination characteristics, accurately identifies the anomalies in the prediction results, avoids generating operation and maintenance decisions based on erroneous prediction results, and improves the fault tolerance of the system.

[0043] Conclusion: The logic verification failed, triggering the parameter feedback analysis mechanism.

[0044] If the logic verification passes, the first prediction result is input into the second prediction model, which is a Transformer encoder model. The second prediction model is used to output a comprehensive health index that represents the overall health status of the system as the second prediction result. The first prediction result and the second prediction result are then fused based on composite operations to generate operation and maintenance guidance decisions. Model structure: Input layer: 32 dimensions (consistent with the input features of the first prediction model); Encoder layer: 2 Transformer encoders, each containing 4 attention heads, with a hidden layer dimension of 128; Output layer: 1-dimensional (comprehensive health index H, value range [0,1]); Activation function: The output layer is normalized to the [0,1] interval using the Sigmoid function.

[0045] The Transformer encoder's attention mechanism can capture long-distance dependencies between operating parameters, providing a more comprehensive assessment of the overall health status of the system compared to traditional models. The design of four attention heads enables multi-dimensional feature extraction, improving the accuracy of the comprehensive health index assessment. The Sigmoid function normalizes the output to the [0,1] interval, intuitively reflecting the system's health level and facilitating the calculation of subsequent operation and maintenance decision indices and threshold comparisons.

[0046] Training process and training data: Same as the first prediction model, with additional labels for system health status (based on manual inspection records and membrane fouling detection results, health status is divided into excellent (H≥0.8), good (0.6≤H<0.8), average (0.4≤H<0.6), and poor (H<0.4)). Optimizer: AdamW optimizer, learning rate = 0.0005, weight decay = 0.001; Loss function: Cross-entropy loss function, validation set accuracy ≥ 92%; Training rounds: 200 rounds, batch size=64, model inference time ≤80ms / inference.

[0047] By labeling health status using manual inspection records and membrane fouling detection results, model training becomes more targeted, and the validation set accuracy of ≥92% ensures the reliability of the comprehensive health index. The weight decay strategy of the AdamW optimizer effectively suppresses model overfitting and improves the model's generalization ability. With an inference time of 80ms / inference and the first prediction model, the overall prediction process takes ≤130ms, meeting the real-time processing requirements of the edge side.

[0048] The fusion of the first and second prediction results based on composite operations is achieved by calculating a multi-dimensional operation and maintenance decision index, specifically including: From the first prediction result, extract the normalized vector ΔV=[δ1,δ2,...,δn] of the short-term change values ​​of the key parameters; The comprehensive health index H is obtained from the second prediction result, wherein a decrease in the H value indicates a deterioration in health status; The operation and maintenance decision index A is calculated using the following composite formula: ; Where, ||ΔV|| represents the norm of the normalized vector of the short-term change values ​​of the key parameters extracted from the first prediction result; The range of H is [0,1], and the lower the H value, the worse the health status. This represents a preset health status threshold, used to divide different decision intervals; The preset weighting coefficients, The preset threshold for the magnitude of change; This is the Sigmoid function, used to provide a smooth decision transition; This formula is a decision function that nonlinearly couples short-term risk (‖ΔV‖) with long-term health status (1-H). Its core innovation lies in determining whether the health status H crosses a threshold. Switch between two risk-weighted strategies to more precisely reflect the operational urgency of the system at different stages of degradation. Health status threshold. The weighting coefficients were determined by analyzing a large amount of historical data, identifying typical values ​​for the health index H corresponding to the transition from mild contamination to the need for chemical cleaning at key points in the system's history. These values ​​were then calibrated using expert experience. , threshold of change Typically, the Sigmoid function is obtained by tuning optimization algorithms (such as grid search or Bayesian optimization) on system simulation models or historical datasets, with the goal of maximizing decision accuracy (e.g., the rate of correctly triggering maintenance). : Used when H≥ In this way, the magnitude of short-term changes is smoothed and compressed to avoid overreaction caused by minor fluctuations when the health condition is still acceptable, thus achieving a smooth transition of decision output.

[0049] At least three ordered decision thresholds are preset: routine monitoring threshold K1, early warning and maintenance threshold K2, and emergency response threshold K3, where K1 < K2 < K3; The calculated operation and maintenance decision index A is compared with the decision threshold, and the corresponding preset decision template is matched according to the interval: if A≤K1, the continuous monitoring decision template is matched, and the generated operation and maintenance guidance decision content is to record data and output the system normal signal according to the first preset cycle. If K1 < A ≤ K2, then the matching optimization and adjustment decision template is used, and the generated operation and maintenance guidance decision content includes adjustment suggestions and target settings for the key parameters that have changed the most in the first prediction result. If K2 < A ≤ K3, then the preventive maintenance decision template is matched, and the generated operation and maintenance guidance decision content includes the recommended maintenance items, maintenance time windows, and risk factors indicated by the second prediction results; If A > K3, then the immediate intervention decision template will be matched, and the generated operation and maintenance guidance decision content will include shutdown inspection instructions, the highest priority alarm, and suspected fault location information.

[0050] Example: Input the first prediction result into the Transformer encoder model, and the output comprehensive health index H=0.75 (in the range of [0.6,0.8], indicating that the system is in good health).

[0051] The operation and maintenance decision index A is calculated by integrating the dual prediction results: The normalized vector ΔV of the first prediction result is extracted: normalized value of the rate of change of permeate flow rate δ1 = 0.21, normalized value of the rate of change of desalination rate δ2 = 0.08, normalized value of the rate of change of inter-stage pressure difference δ3 = 0.53, i.e., ΔV = [0.21, 0.08, 0.53]; the L2 norm of ΔV is calculated: ||ΔV|| = √(0.21). 2 +0.08 2 +0.53 2 =0.576; Since H=0.75≥ =0.6, calculate A using the second branch formula; =0.576 - 0.3 = 0.276; (0.276) = 0.569; Finally, A = 0.569 × (1 - 0.75) = 0.142; The comprehensive health index H=0.75 accurately reflects the system's current good health status, consistent with actual operation; the normalized vector ΔV quantifies the short-term changes in key parameters, and the L2 norm comprehensively reflects the overall trend of multiple parameters; the composite calculation formula uses the health status threshold... The decision-making branches are divided and the Sigmoid function is used to achieve a smooth transition, avoiding abrupt changes in the decision-making logic. The final calculated operation and maintenance decision index A=0.142 accurately integrates two types of information: short-term parameter changes and overall health status, providing a quantitative basis for hierarchical operation and maintenance decisions.

[0052] Comparing A=0.142 with the preset decision threshold: A=0.142≤K1=0.2, matching the continuous monitoring decision template. The generated operation and maintenance guidance decision content is as follows: Record system operation data according to the first preset cycle (1 hour), including key parameters such as influent flow rate, product flow rate, desalination rate, and inter-stage pressure difference; output a normal system signal (indicator light is constantly green, no alarm information); background log records: The system is currently operating in a steady state, key parameters show stable short-term changes, and the overall health status is good. It is recommended to maintain the current operating parameters and continue monitoring.

[0053] Based on the quantitative operation and maintenance decision index A, corresponding decision templates are matched to achieve precise hierarchical operation and maintenance decisions. Continuous monitoring of the decision template output ensures the tracking of the system's normal operating status while avoiding unnecessary operation and maintenance interventions, thus reducing operation and maintenance costs. Log recording provides a basis for subsequent system operation status tracing, facilitating operation and maintenance personnel to analyze the system's long-term operating trends. If the logical verification fails, a parameter feedback analysis mechanism is triggered to locate abnormal operating parameters. The trigger parameter feedback analysis mechanism is used to locate abnormal operating parameters, including: For each operating parameter, calculate the average rate of change of the cumulative physical quantity of that operating parameter within a preset time window relative to the window length, as the cumulative change trend of that operating parameter; The average rate of change is compared with the baseline range of the corresponding parameter established based on historical normal operation data. If it exceeds the baseline range, the operating parameter is determined to be an abnormal operating parameter.

[0054] Then, replace the abnormal operating parameters with historical statistical values ​​and re-verify the logic. If the logic verification after substitution passes, the device fault diagnosis result for the abnormal operating parameters will be output.

[0055] Example: Calculate the average rate of change of all operating parameters within a preset time window (i.e., 30 minutes): Average rate of change of influent conductivity = (1400-850) μS / cm / 30 minutes = 18.33 μS / cm·min; The baseline range for the average rate of change of influent conductivity, established based on historical normal data, is [0.1, 2.5] μS / cm·min. The current value of 18.33 μS / cm·min exceeds the baseline range, and the influent conductivity is determined to be an abnormal operating parameter.

[0056] By calculating the average rate of change within a preset time window, rather than the parameter value at a single moment, normal fluctuations and abnormal abrupt changes can be effectively distinguished, avoiding misjudgments. Establishing a benchmark range based on historical normal data makes the anomaly judgment criteria more aligned with the system's own operating patterns, which is more targeted than fixed thresholds, enabling rapid and accurate location of abnormal operating parameters. Using historical statistical values ​​of influent conductivity (average of 820 μS / cm during normal operation over the past 3 months) to replace outliers, a new 32-dimensional operating parameter dataset is generated, and further validation using the third set of verification rules is performed. Verification Rule 2: The average conductivity of the influent is 820 μS / cm, and the predicted direction of change in the conductivity of the product water is increasing, which is consistent with the theoretical response direction of membrane desalination characteristics, thus satisfying the rule; All other rules were met, and the re-verification passed.

[0057] By using historical statistical values ​​of abnormal parameters as substitutes, the continuity of the subsequent verification process was ensured, and the interference of abnormal parameters on the verification results was avoided. The successful re-verification proved that the predicted results of other parameters and the system status were normal except for the influent conductivity, which provided a clear direction for fault diagnosis and improved the robustness of the method.

[0058] Output device fault diagnosis result: The influent conductivity suddenly increased abnormally (current value 1400 μS / cm, exceeding the normal range), suspected to be raw water pollution or conductivity sensor malfunction. Recommendations: 1. Test the raw water source quality; 2. Calibrate the influent conductivity sensor; 3. Closely monitor the product water quality. If the product water conductivity continues to rise to ≥30 μS / cm, the emergency response procedure must be initiated. The fault diagnosis result clearly identifies the abnormal parameters and suspected causes, providing maintenance personnel with precise troubleshooting directions and avoiding blind repairs. The targeted recommendations (testing raw water quality, calibrating the sensor) are highly operable, helping maintenance personnel quickly handle the fault and reduce its impact on system operation. The emergency response procedure further reduces risk, demonstrating the comprehensiveness and practicality of the fault diagnosis.

[0059] If the calculated operation and maintenance decision index A = 0.35 at a certain moment (K1 = 0.2 < A ≤ K2 = 0.5), match the optimization and adjustment decision template: The key parameter with the largest change in the first prediction result is the inter-segment pressure difference (change rate +8.7%). Operation and maintenance guidance decision: It is recommended to adjust the inlet water flow rate to 46m³. 3 / h (target setpoint) to reduce the operating load on the membrane module; control the influent pressure to stabilize at 1.15MPa to slow down the growth rate of inter-section pressure difference; increase the frequency of inter-section pressure difference monitoring by 1 time per day to continuously track the fouling trend.

[0060] The early warning maintenance decision provides specific adjustment suggestions and target settings for the key parameter (inter-segment pressure difference) with the greatest changes. It is highly operable and can intervene in the early stage of problems, slow down the rate of membrane fouling, and extend the service life of membrane elements. The suggestion to increase the monitoring frequency ensures continuous tracking of risk trends, prevents problems from escalating, and achieves the core goal of preventive maintenance.

[0061] If the operation and maintenance decision index A = 0.92 (A > K3 = 0.8) is calculated at a certain moment, the immediate intervention decision template is matched: Operational guidance decision generation: Emergency alarm! System overall health index H=0.35 (poor), inter-section pressure difference prediction change rate +15.2% (expected to rise to 0.52MPa within 1 hour, approaching the safety limit), suspected severe membrane element blockage. Immediately shut down the system for inspection, focusing on the fouling of membrane modules 1-3, and perform chemical cleaning; record the pressure difference of each section and the appearance of the membrane elements during the shutdown. After the inspection is completed, restart the system and continuously monitor for 1 hour.

[0062] Emergency response decisions, through the highest priority alarms and clear shutdown instructions, quickly respond to serious fault risks, avoid irreversible damage to membrane elements due to excessive contamination, and reduce production interruption losses; the prompts for suspected fault location and key points of investigation shorten fault handling time and improve operation and maintenance efficiency.

[0063] The predictive model of this invention supports online iterative optimization: edge computing nodes periodically (ideally once a month) collect newly added normal operation data and fault handling data from the system, and incrementally train the first and second predictive models to update model parameters. During incremental training, a hybrid dataset of new data and a subset of historical training data is used (30% new data and 70% historical data subset) to avoid the model forgetting historical features; the training rounds are 50-100, and the learning rate is reduced to 1 / 5 of the initial value to ensure smooth updating of model parameters. Online iterative optimization enables the model to adapt to performance degradation and changes in operating conditions during long-term system operation, maintaining high predictive accuracy; the incremental training mode reduces computational resource consumption, adapts to edge deployment requirements, and extends the effective lifespan of the model.

[0064] As one embodiment, if the logic verification still fails after replacement, the parameter feedback analysis mechanism is re-triggered, and a secondary fault location is executed: The preset time window is shortened to a second duration, and the average rate of change of all operating parameters within the shortened window is calculated. Simultaneously, collect and analyze the system's hardware status data; Based on the average rate of change and the hardware status data, the fault type is determined to be either a multi-parameter collaborative anomaly or a hardware fault, and the corresponding secondary fault diagnosis result is output.

[0065] Taking a reverse osmosis system (3 sets of BW30-4040 membrane modules, designed product water flow rate of 50 m³ / h) in an industrial water treatment plant as an example, the original preset time window was 30 minutes. After the initial verification failed, the influent conductivity was identified as an abnormal parameter. After replacing it with a historical statistical value (820 μS / cm), the re-verification still failed, and a secondary fault location was performed. Shorten the time window: Reduce the time window from 30 minutes to 15 minutes and calculate the short-term average rate of change of all operating parameters. The rate of change of permeate flow rate is -7.8% / 15min (baseline range [-2.0%, +3.0%]), the rate of change of desalination rate is -4.9% / 15min (baseline range [-1.0%, +0.5%]), and the rate of change of inter-stage pressure difference is +11.2% / 15min (baseline range [-0.8%, +2.5%]), all of which exceed the baseline range. Hardware status data acquisition: All sensors (flow rate, pressure, conductivity) show normal communication connection, the PLC operation log has no error codes, valve execution feedback is in place, the membrane module pressure signal is stable, and there are no hardware abnormality prompts; Fault Judgment and Diagnosis Output: Because the short-term change rates of the three key parameters all exceed the range and the hardware status is normal, it is determined to be a multi-parameter coordinated anomaly, and a secondary fault diagnosis result is output: Level 2 Fault: Multi-parameter coordinated anomaly. Suspected cause: Sudden pollution of raw water quality (such as excessive suspended solids or organic matter) leading to a sharp drop in membrane module filtration performance. Recommended actions: 1. Immediately sample and test the raw water for COD and suspended solids content; 2. Suspend system operation and chemically clean the membrane modules; 3. After cleaning, restart the system and monitor key parameters at a frequency of once every 5 minutes. Resume normal operation after confirming that the system has returned to normal.

[0066] When the logical verification of the prediction result fails, the initial assumption is that a single data source is distorted, causing the inconsistency. A parameter feedback analysis mechanism is used to locate the most suspicious parameter, and its historical normal value is used to replace it to test whether the prediction is reasonable if this parameter is normal. If the hypothesis is true and re-verification passes, the problem converges to a single parameter anomaly. If the hypothesis is refuted (re-verification still fails), a second-level final diagnosis is initiated, indicating that the fault may involve the synergistic effect of multiple parameters or a fundamental system problem.

[0067] At this point, a crucial shift occurs in the diagnostic dimensions: on the time scale, the analysis window is shortened to capture short-term mutations and collaborative patterns of parameters; on the data scale, hardware status data is introduced for cross-domain fusion analysis. Ultimately, by comprehensively judging short-term collaborative change patterns and hardware health status, a fundamental distinction is made between real complex process faults and underlying hardware / data acquisition faults, completing intelligent reasoning from anomalies to root causes. This achieves gradient-based fault identification, clearly distinguishing between single-point sensor distortion, multi-parameter correlated process faults, and data acquisition system hardware faults, fundamentally changing the traditional crude mode of alarming upon anomalies in monitoring systems and greatly reducing false alarms. Simultaneously, through parameter replacement verification in the first-level diagnosis, it can tolerate and identify anomalies from a single data source, ensuring continuous and stable operation under imperfect data conditions; and when deeper contradictions arise, its ability to initiate ultimate diagnosis demonstrates the system's metacognitive understanding of its own judgment uncertainties and its intelligent proactive in-depth investigation.

[0068] It should be noted that the calculation formulas and all parameters involved in the calculations in this invention have been dimensionless beforehand. The process of dimensionless processing is well known in the industry and will not be described here.

[0069] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A smart operation and maintenance method for industrial reverse osmosis systems based on edge computing, characterized in that, include: The operating data of the reverse osmosis system is collected and preprocessed in real time at the edge computing node, and time-domain and frequency-domain features are extracted to form an operating parameter dataset. The operating parameter dataset is input into the first prediction model deployed on the edge side to predict the short-term changes of key operating parameters within a preset time period, which is used as the first prediction result. Key operating parameters include at least the permeate flow rate, desalination rate, and inter-stage pressure differential; The first prediction result is logically verified, and the verification mode used in the logical verification is dynamically selected according to the real-time operating conditions of the system. If the logic verification passes, the first prediction result is input into the second prediction model to evaluate the system's comprehensive health index as the second prediction result. The first prediction result and the second prediction result are then fused based on composite operations to generate operation and maintenance guidance decisions. If the logic verification fails, the parameter feedback analysis mechanism is triggered to locate the abnormal operating parameters. The historical statistical values ​​of the abnormal operating parameters are then used as substitutes before the logic verification is performed again. If the logic verification after substitution passes, the device fault diagnosis result for the abnormal operating parameters will be output.

2. The intelligent operation and maintenance method for industrial reverse osmosis systems based on edge computing according to claim 1, characterized in that, The trigger parameter feedback analysis mechanism is used to locate abnormal operating parameters, including: For each operating parameter, calculate the average rate of change of the cumulative physical quantity of that operating parameter within a preset time window relative to the window length, as the cumulative change trend of that operating parameter; The average rate of change is compared with the baseline range of the corresponding parameter established based on historical normal operation data. If it exceeds the baseline range, the operating parameter is determined to be an abnormal operating parameter.

3. The intelligent operation and maintenance method for industrial reverse osmosis systems based on edge computing according to claim 2, characterized in that, The verification mode is dynamically selected based on the real-time operating conditions of the system, including: Based on the stability of key operating parameters, select the first set of verification rules for steady-state verification; Based on the gradual trend of system performance indicators or a preset period, select a second set of verification rules for in-depth performance verification. Based on the sudden changes or exceeding of limits in the operating parameters, a third set of verification rules is selected for anomaly and boundary checks.

4. The intelligent operation and maintenance method for industrial reverse osmosis systems based on edge computing according to claim 3, characterized in that, The verification of the first set of verification rules simultaneously satisfies: The difference between the real-time collected influent flow rate and the sum of the product water flow rate and concentrate flow rate is less than the first allowable error threshold. The real-time ratio of inlet water pressure to product water flow rate is within the normal reference range determined based on historical steady-state data of the system.

5. The intelligent operation and maintenance method for industrial reverse osmosis systems based on edge computing according to claim 3, characterized in that, The validation of the second set of validation rules simultaneously satisfies: The system recovery rate is based on real-time traffic calculation, and the standard deviation of the system recovery rate is less than the preset stability threshold. The absolute deviation between the system desalination rate predicted based on operating parameters and the desalination rate calculated based on real-time water quality measurements is less than a preset consistency threshold. The rate of change of the moving average of the inter-segment pressure difference within the preset time period is less than the preset pollution growth threshold.

6. The intelligent operation and maintenance method for industrial reverse osmosis systems based on edge computing according to claim 3, characterized in that, The verification of the third verification rule set simultaneously satisfies: The real-time calculated system recovery rate is ≤ the maximum allowable recovery rate specified for the membrane element, and the real-time inlet water pressure is ≤ the system safety pressure limit; When the change in influent conductivity exceeds the abrupt change threshold, the predicted change in product water conductivity is consistent with the theoretical response direction determined based on membrane desalination characteristics. After a step change in the influent flow rate, the response time for the predicted product flow rate to reach a steady state is within the theoretical time range calculated based on the system's hydraulic volume.

7. The intelligent operation and maintenance method for industrial reverse osmosis systems based on edge computing according to claim 1, characterized in that, The first prediction model is a lightweight long short-term memory network model; the second prediction model is a Transformer encoder model, used to output a comprehensive health index that characterizes the overall health status of the system.

8. The intelligent operation and maintenance method for industrial reverse osmosis systems based on edge computing according to claim 1, characterized in that, The fusion of the first and second prediction results based on composite operations is achieved by calculating a multi-dimensional operation and maintenance decision index, specifically including: From the first prediction result, extract the normalized vector ΔV=[δ1,δ2,...,δn] of the short-term change values ​​of the key parameters; The comprehensive health index H is obtained from the second prediction result, wherein a decrease in the H value indicates a deterioration in health status; The operation and maintenance decision index A is calculated using the following composite formula: ; Where |ΔV| represents the norm of the normalized vector of the short-term change values ​​of the key parameters extracted from the first prediction result; the range of H is [0,1], and the lower the value of H, the worse the health status; This represents a preset health status threshold, used to divide different decision intervals; The preset weighting coefficients, The preset threshold for the magnitude of change; This is the Sigmoid function, used to provide a smooth decision transition; The operation and maintenance decision index A is compared with a preset decision threshold to generate a corresponding operation and maintenance guidance decision.

9. The intelligent operation and maintenance method for industrial reverse osmosis systems based on edge computing according to claim 8, characterized in that, The specific components of generating operation and maintenance guidance decisions include: At least three ordered decision thresholds are preset: routine monitoring threshold K1, early warning and maintenance threshold K2, and emergency response threshold K3, where K1 < K2 < K3; The calculated operation and maintenance decision index A is compared with the decision threshold, and a corresponding preset decision template is matched according to the interval: If A≤K1, then the continuous monitoring decision template is matched; If K1 < A ≤ K2, then match and optimize the decision template. If K2 < A ≤ K3, then the preventive maintenance decision template is matched; If A > K3, then the immediate intervention decision template is matched.

10. The intelligent operation and maintenance method for industrial reverse osmosis systems based on edge computing according to claim 1, characterized in that, If the logic verification still fails after replacement, the parameter feedback analysis mechanism will be retried, and a secondary fault location will be executed: The preset time window is shortened to a second duration, and the average rate of change of all operating parameters within the shortened window is calculated. Simultaneously, collect and analyze the system's hardware status data; Based on the average rate of change and the hardware status data, the fault type is determined to be either a multi-parameter collaborative anomaly or a hardware fault, and the corresponding secondary fault diagnosis result is output.

Images